Finetuning Generative Large Language Models with Discrimination Instructions for Knowledge Graph Completion